Claude formalizes Fermat's Last Theorem in 11 days: 13M lines of Lean, 5x Mathlib

量子位 · wechat · 2026-09-05

Anthropic says Claude completed the first end-to-end, computer-checkable formalization of Fermat's Last Theorem: 13 million lines of Lean code with over 30,000 intermediate theorems (29,500 used), more than 5x the size of Mathlib. It didn't find a new proof — it translated Wiles' 1994 human proof into a fully machine-checkable form, a project the math community had planned as a multi-year effort, finished by Claude in 11 days.

Key details

Led by Anthropic researcher Tianyi Peng (Columbia assistant professor, Tsinghua Yao class, MIT PhD), who initially only intended to see how far the project could be pushed.
Multiple agents worked in parallel on definitions, lemmas, higher-level theorems, and final assembly; 6 billion output tokens, with the internal model roughly comparable to Claude Fable 5.1.
Early multi-agent collaboration broke down — failed attempts made up only 7% of final code — until the team adopted Prove2Me, a platform that decomposes the proof into a DAG of theorem nodes, manages dependencies, speeds up Lean compilation, and stores natural-language descriptions for retrieval.
Human mathematical guidance was minimal; the final proof relies only on Lean's three standard axioms, and programmatic comparison confirmed the theorem statement matches Mathlib's. Lean, not the model, served as judge.
Reviewer Kevin Buzzard (Imperial College) called it "an extraordinary result in automated formalization," suggesting large-scale formalization of mathematical literature may finally be able to accelerate.

Meanwhile, OpenAI began rolling GPT-6 Astra out to Pro/Business/Enterprise users, emphasizing long-horizon agentic tasks, software engineering, computer use, and scientific work.

Original post →

Claude formalizes Fermat's Last Theorem in 11 days: 13M lines of Lean, 5x Mathlib

More from coding & agent